SMAC for automatic algorithm selection and hyperparameter optimization of Scikit-learn machine learning pipelines. Inspired and based on Auto-sklearn. This version incorporates pipeline caching to speed up optimization.
Use python 3.6.
We recommend using an anaconda environment. Some commands on this page only work in an anconda environment. Therefore, first download and install anaconda or miniconda. Then create an environment called pcsmac with following command
conda create -n pcsmac python=3.6
Install SMAC3 with pipeline caching integration from command line. Install SWIG for the C++ random forest.
conda install swig
git clone https://github.com/jtuyls/SMAC3_4_PC.git
cd SMAC3_4_PC
cat requirements.txt | xargs -n 1 -L 1 pip install
python setup.py install
Clone repository and install requirements.
pip install -r requirements.txt
For installation on Mac OS X there is a problem with installing SMAC3 as above. SMAC is dependenct on pyrfr that uses the gcc compiler for c++ while Mac OS X uses the clang compiler. Therefore we install a gcc compiler within anaconda. Furthermore, install SWIG for the C++ random forest.
Install SMAC3 with pipeline caching integration from commandline in anaconda environment. Therefore first install pyrfr separately with anaconda gcc compiler.
conda install swig
conda install gcc
CC=/Users/[username]/anaconda/bin/gcc pip install pyrfr --no-cache-dir
git clone https://github.com/jtuyls/SMAC3_4_PC.git
cd SMAC3_4_PC
cat requirements.txt | xargs -n 1 -L 1 pip install
python setup.py install
Clone repository and install requirements.
pip install -r requirements.txt
First go to the directory above pc_smac. Then the optimization can be started using following generic command:
python -m pc_smac.run -a=[acquistion function, STRING: e.g pc-m-pceips]
-di=[double intensification enabled: INT: 1 (yes) 0 (no)]
-w=[wallclock time in seconds, INT: e.g. 1800]
-l=[Location of the dataset: STRING: DEFAULT: if not provided, data/46_bac is used as example]
-m=[OPTIONAL: memory limit in Mb: INT: default=6000]
-r=[OPTIONAL: maximum number of runs: INT: DEFAULT=10000]
-c=[OPTIONAL: cutoff for each evaluation in seconds: INT: DEFAULT=1800]
-s=[OPTIONAL: stamp used for identifying the directory and names of the results: STRING: DEFAULT=stamp]
-o=[OPTIONAL: output directory: STRING: DEFAULT=pc_smac/output]
-cd=[OPTIONAL: cache directory: STRING: DEFAULT=pc_smac/cache]
-ps=[OPTIONAL: pipeline space used: STRING: DEFAULT=None; Example of small space: nystroem_sampler-sgd]
-ifs=[OPTIONAL: intensification fold size: INT: DEFAULT=10]
-sn=[OPTIONAL: splitting number ! only for sigmoid and mrs random search version: INT: DEFAULT=5]
-rs=[OPTIONAL: random splitting enabled ! only for mrs random search version: INT: DEFAULT=0 (no)]
Execute following command to test optimization with pipeline caching (and marginalization) for 1800 seconds or 100 evaluations on OpenML dataset 46.
python -m pc_smac.run -a=pc-m-pceips -di=0 -w=1800 -r=100 -c=100 -m=4000 -s=pc-m-pceips
Don't forget to clean caches afterwards because they can take quite some space!
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent